Lesson objective
Review a proposed construction automation for excessive functionality, excessive permissions and excessive autonomy, and specify the design change that removes each.
Plain-language explanation
Why this matters now
When an automation does damage, the investigation rarely finds that the model was malicious. It finds that the tool could do more than the task needed, reach more than the task needed, or do it with less checking than the task needed. This lesson names those three root causes and turns each into a design question you ask before anything is connected.
The security community has been moving the same way.
- Professional guidance (OWASP GenAI Security Project, OWASP Top 10 for LLM Applications 2026, PDF dated 4 August 2026; voluntary): Excessive Agency is entry LLM03:2026. The project states it climbed to third place from the 2025 edition. The same document maps Excessive Agency, in agentic systems, to three entries of the separate OWASP Top 10 for Agentic Applications for 2026: ASI02 Tool Misuse, ASI03 Identity & Privilege Abuse and ASI08 Cascading Failures. Both lists are community documents, not regulations or certifiable standards.
- The project leads' letter in the 2026 list gives the design stance in one line: "Stop trying to build a model that cannot be fooled. Build the system around it, so that when the model is fooled, and it will be, nothing important breaks."
- On misinformation, the leads note that voters placed it near the bottom while the incident record placed it near the top: "When a model's fluent, confident output drives a decision or a tool call, a wrong answer turns into a wrong action."
That last sentence is the heart of Level 5. A wrong answer in a chat window costs a person some time. A wrong answer wired to a tool changes a log, sends a letter or releases a document.
A question to sit with
If the tool you are about to connect misread one document tomorrow, what is the worst thing it could do with the access you are about to give it, and did anyone need it to be able to do that?
What goes wrong
This programme names three root causes to design against. The framing is an AI Coalition Network recommendation for review purposes; it is not a definition taken from OWASP, and this lesson does not paraphrase the definitions of any OWASP entry.
Excessive functionality. The automation carries tools it does not need. A mailbox connector installed to read incoming submittals can also send, forward and delete. A document plug-in added to summarise specifications can also overwrite files. Each extra tool is an action the automation could take on a bad day.
Excessive permissions. The tools it needs are connected with more reach than the task needs. An RFI logging automation for one project runs under an account that can see every project, the accounting system and the company drive. A read task runs with write access because "it was easier to set up". CISA's 2026 announcement names privilege creep as a risk: access that grows quietly and never shrinks.
Excessive autonomy. The automation takes consequential steps without a person seeing them first. It posts, sends or changes status on its own because the gate was inconvenient or never built.
The three compound. A tool with one extra function, a little extra reach and no gate can turn one misread document into a chain of wrong actions across several projects. That chaining is why OWASP's mapping points Excessive Agency at cascading failures as well as at tool misuse and privilege abuse.
The practical answer
Run a three-question design review before connection. For each proposed automation, answer the three questions in writing. The review below is this programme's design.
| Root cause | The review question | Construction example | The design answer |
|---|---|---|---|
| Excessive functionality | Which tools does the task need, and which tools does the connector carry? | Mailbox connector for submittal intake can send, forward and delete | Use a read-only connector, or disable every function beyond read; list the permitted tools by name |
| Excessive permissions | What is the smallest set of projects, folders and records the task needs? | Logging automation runs under an admin account across all projects | A dedicated automation identity scoped to one project's submittal folder and log |
| Excessive autonomy | Which steps leave the company or cannot be undone, and does a person see each first? | Automation posts "approved" status to a shared project system | The automation proposes the status; a named person posts it |
Start from zero and add. Build the permission list the way you would build a subcontract scope: start from nothing and add each item the work requires, rather than starting from "everything" and removing what looks dangerous. A tool or permission that nobody can justify in one sentence comes off.
Separate reading from writing. Many construction automations only need to read: summarise daily reports, compare a submittal against a specification section, list overdue RFIs. Give them read access only. Where a write is needed, make it the narrowest write that does the job: append a row to a log, not edit any record in the system. Module 05 develops this as "read first, write with approval."
Scope to the project, not the company. A contracting business works project by project, and owners expect their project information to stay on their project. An automation built for one job should reach that job's records only. When it is extended to another job, that is a register change with a named approver, not a checkbox. Module 08 builds automation identities and project segregation properly.
Put the gate where the consequence is. Excessive autonomy is fixed by placing a person at the point where the step becomes consequential, not by reviewing everything. A rung 2 filing step inside the company may run on its own. The step that leaves the company does not.
Assume the model will be fooled. Following the OWASP leads' stance, design as if the model will sometimes misread a document or follow an instruction hidden in one. Ask what the automation could do in that moment. If the answer is "not much, and a person would see it before it left the company", the design is sound. If the answer is "send, delete or change something across projects", the design is not. Module 10 treats injected instructions in submittals, invoices and inbound mail from the defender's side.
Why "the model is usually right" is not a design argument
An automation runs many times a week across many projects. A tool that is right almost every time will still be wrong on some run, and the design decides what that wrong run can touch. Accuracy lowers the frequency of bad outcomes; scope and gates limit their size. Both matter, and only the second is fully in the company's control.
Record the review. Keep the three answers with the automation's register entry, dated, with the reviewer's name. When someone later asks for "just one more permission", the review shows what was decided and why, and the request becomes a change to a written decision.
Remember this
- Three root causes to design against: excessive functionality, excessive permissions and excessive autonomy.
- OWASP's 2026 LLM list ranks Excessive Agency third (LLM03:2026) and maps it to ASI02, ASI03 and ASI08; both lists are voluntary community documents.
- Build permissions up from zero, separate reading from writing, and scope to the project.
- Put the human gate where the step becomes consequential, especially where it leaves the company.
- Design as if the model will be fooled, so that when it is, nothing important breaks.
Construction example
Bellhaven Mechanical is a fictional 40-person mechanical subcontractor that also acts as a prime on small design-build jobs. Its operations manager, Priya Castellanos, asks the company's outside integrator to connect an AI assistant that will read the shared project inbox, pick out submittal returns from engineers and update the submittal log. The integrator's first proposal uses the connector the assistant ships with, signed in under Priya's own account. Priya's technology lead, Owen Tsai, runs the three-question review. On functionality, the connector can read, send, forward, delete and move mail, and can edit any spreadsheet the account can reach; the task needs to read mail and append rows to one log. On permissions, Priya's account reaches all 23 active projects, the estimating folder and the accounting export; the task needs one project's submittal mail and one log. On autonomy, the proposal lets the assistant mark a submittal "approved as noted" in the log that the general contractor's project manager can see. Owen writes up the answers. The redesign uses a dedicated automation account, not Priya's, with read access to a single project mailbox folder and permission to append rows to that project's log. Sending, forwarding and deleting are disabled at the connector. The assistant proposes each status change in a separate review column, and the project engineer copies the status into the shared log after checking the returned stamp against the engineer's own marks. Owen records the review in the register entry with the date and his name. When the integrator later suggests "turning on the other projects while we are at it", Priya treats it as a register change, and asks who will own each additional project's log before anything is added.
Responsible-use warning
Do not connect an automation under a person's own account because it is quicker. A person's account carries every project, folder and system that person can reach, so the automation inherits all of it, and its actions cannot be told apart from the person's in the records. That is excessive permissions and a broken trail in one step. Give each automation its own identity with the narrowest reach the task needs, and never combine write access, reach across projects and no human gate in one automation.
Human-review requirement
The automation owner runs the three-question review and records it. The company's information security lead or technology owner confirms the tool list and the scope were actually applied in the connector, not just written down. The person who owns the affected output checks the gate position: the project manager for project records and correspondence, the qualifying individual or equivalent responsible person for anything bearing on construction operations. Each signs off by name before the automation touches a live project.
Practical exercise
Use a fictional automation, or one proposed in your own company with nothing confidential.
- Describe the automation in two sentences: its trigger and the result it should produce.
- List every tool the connector or platform makes available to it, including those nobody plans to use.
- Mark each tool "needed" or "not needed", completing the sentence "without it, the automation cannot ___" for each needed one.
- List every project, folder, mailbox and system the account it will run under can reach.
- Write the smallest scope the task needs, naming one project where possible.
- Identify every step that leaves the company or cannot be undone, and state who sees each step before it happens.
- Check whether the design combines write access, reach across projects and no human gate. If it does, remove one and explain which.
- Write the review up as three short paragraphs, date it, name the reviewer, and note the connector settings that must change to match it.
- Assume the model misreads one document on its next run. Describe the worst action it could take under your redesigned scope, and compare it with the worst action under the original design.
- List who must confirm the connector settings were actually changed, not just written down, and when the scope will next be re-checked against the live configuration.
Knowledge check
-
What are the three root causes this lesson designs against?
Answer guidance: Excessive functionality, where the automation carries tools it does not need; excessive permissions, where needed tools reach more than the task requires; and excessive autonomy, where consequential steps happen without a person seeing them first.
-
Where does Excessive Agency sit in the OWASP LLM Top 10 2026, and to which agentic entries is it mapped?
Answer guidance: It is LLM03:2026, having climbed to third. The document maps it in agentic systems to ASI02 Tool Misuse, ASI03 Identity & Privilege Abuse and ASI08 Cascading Failures. Both OWASP lists are voluntary community documents.
-
Which combination does this programme recommend refusing outright?
Answer guidance: Write access, reach across more than one project and no human gate in the same automation. Removing any one of the three sharply reduces the worst case of a misread document.
-
Why is running an automation under a person's own account a design defect?
Answer guidance: The automation inherits every project and system that person can reach, which is excessive permissions, and its actions cannot be told apart from the person's in the records, which breaks the trail.
Sources and further reading
- OWASP Foundation, OWASP GenAI Security Project — OWASP Top 10 for LLM Applications 2026 (2026-08-04). Accessed 2026-09-15.
- OWASP Foundation, OWASP GenAI Security Project — OWASP Top 10 for Agentic Applications for 2026 (2025-12-09). Accessed 2026-09-15.
- Cybersecurity and Infrastructure Security Agency (CISA), Australian Signals Directorate's Australian Cyber Security Centre, and U.S. and international partners — Careful Adoption of Agentic Artificial Intelligence (AI) Services (2026-05-01). Accessed 2026-09-15.
- National Institute of Standards and Technology — Artificial Intelligence Risk Management Framework (AI RMF 1.0), NIST AI 100-1 (2023-01-26). Accessed 2026-09-15.